The idea the LLMs would work in bytecode kind of ignores why they are good at higher level code, IMO. Bytecode would fill and overflow context so fast and locality of reasoning is much harder.